Journal Pre-proofs

Recognition and separation of fringe patterns in deflectometric measurement

of transparent elements based on empirical curvelet transform

Peide Yang, Ting Chen, Dongfang Wang, Lu Ye, Yunuo Chen, Wei Lang,
Xiangchao Zhang

PII: S0263-2241(24)01127-8
DOI: https://doi.org/10.1016/j.measurement.2024.115242
Reference: MEASUR 115242

To appear in: Measurement

Received Date: 24 March 2024

Revised Date: 3 July 2024
Accepted Date: 4 July 2024

Please cite this article as: P. Yang, T. Chen, D. Wang, L. Ye, Y. Chen, W. Lang, X. Zhang, Recognition and
separation of fringe patterns in deflectometric measurement of transparent elements based on empirical curvelet
transform, Measurement (2024), doi: https://doi.org/10.1016/j.measurement.2024.115242

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover
page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version
will undergo additional copyediting, typesetting and review before it is published in its final form, but we are
providing this version to give early visibility of the article. Please note that, during the production process, errors
may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar
technologies.
 Recognition and separation of fringe patterns in deflectometric measurement of

transparent elements based on empirical curvelet transform

Recognition and separation of fringe patterns in

deflectometric measurement of transparent elements
based on empirical curvelet transform
Peide Yang1, Ting Chen2, Dongfang Wang3, Lu Ye1,3, Yunuo Chen1, Wei Lang1, Xiangchao

Zhang1*

1. Shanghai Engineering Research Center of Ultra-Precision Optical Manufacturing, School of

Information Science and Technology, Fudan University, Shanghai 200438, China
2. Academy for Engineering and Technology, Fudan University, Shanghai, 200433, China
3. Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai, 200083,
PR China
* Corresponding author. Email: [email protected].

Abstract

In the deflectometric measurement of transparent elements, it is critical to separate the superposed

fringes reflected from the front and rear surfaces to establish the pixel correspondences between the

camera and screen pixels. But it is a challenging task due to the low reflectivity and uneven

background of transparent elements. A reliable fringe separation method is proposed based on the

empirical curvelet transform. The captured fringe images are decomposed into different modes

according to their directions, periods, curvatures or modulation coefficients, and then a weighted

permutation entropy-based fusing algorithm is developed to automatically aggregate those over-

separated patterns. By employing the proposed method, the measurement accuracy is significantly

improved, the number of demanded images is effectively reduced, and high-precision, automatic,

and rapid measurement is thereby achieved. The measurement accuracy of spherical and cylindrical

lenses can achieve 73 nm and 332 nm Root-of-Mean-Squares-Error, respectively.

CRediT authorship contribution statement

 Peide Yang: Methodology, Writing - original draft. Ting Chen:

Investigation. Dongfang Wang: Validation. Lu Ye: Writing - review &

editing. Yunuo Chen: Investigation. Wei Lang: Investigation. Xiangchao

Zhang: Supervision, Writing - review & editing, Conceptualization,

Funding acquisition.

1. Fringe superposition is an essential challenge for the deflectometric measurement of transparent

elements.
2. A general-purpose fringe separation strategy is proposed by the empirical curvelet based
frequency domain filtering.
3. Over-separated patterns are recombined based on the weighted permutation entropy method to
avoid phase distortion.

Abstract

In the deflectometric measurement of transparent elements, it is critical to separate the superposed

fringes reflected from the front and rear surfaces to establish the pixel correspondences between the

camera and screen pixels. But it is a challenging task due to the low reflectivity and uneven

background of transparent elements. A reliable fringe separation method is proposed based on the

empirical curvelet transform. The captured fringe images are decomposed into different modes

according to their directions, periods, curvatures or modulation coefficients, and then a weighted

permutation entropy-based fusing algorithm is developed to automatically aggregate those over-

separated patterns. By employing the proposed method, the measurement accuracy is significantly

improved, the number of demanded images is effectively reduced, and high-precision, automatic,

and rapid measurement is thereby achieved. The measurement accuracy of spherical and cylindrical

lenses can achieve 73 nm and 332 nm Root-of-Mean-Squares-Error, respectively.

Keywords: transparent elements; deflectometry; fringe pattern separation; empirical curvelet

transform.
1. Introduction

Transparent optical elements, including lenses, prisms, and others, play a crucial role in optical

systems such as high-end cameras, VR/AR devices, and lighting equipment. They find extensive

applications across various industries, including automotive, electronics, aerospace, and medical

equipment [1, 2]. The form accuracy of these transparent elements directly affects the performance

of the optical systems, e.g., the form error of an injection molding lens is normally required to be

from hundreds of nanometers to one micron [3].

Various methods have been developed for measuring the form quality of complex objects.

Traditional contact measurement methods, such as the coordinate measuring machine, scanning

profilometer, and other instruments, are inefficient and prone to damaging the surfaces under test,

which are not preferred in practice [4]. Subsequently, various non-contact optical measurement

methods have been proposed, which possess advantages such as non-contact, full-field measurement,

high accuracy, and high efficiency. Structural light profilometry including the Moiré profilometry,

the Fourier transform profilometry, and modulation measuring profilometry is widely employed.

The Moiré profilometry allows the reconstruction of a three-dimensional (3D) surface by analyzing

the Moiré patterns generated by two gratings. The contours of the measured objects can be specified

by analyzing the Moiré fringes [5]. The Fourier transform profilometry obtains the height map of

an object through the Fourier transform of the captured fringe patterns. It offers higher sensitivity

than traditional Moiré methods and reduces errors from higher-order harmonic distortions in the

gratings [6, 7]. The modulation measuring profilometry analyzes the modulation of the captured

sinusoidal fringes, and this technology normally applies to coaxial systems in which the projection

axis coincides with the observation axis. It is capable of measuring objects with significant height

variations or discontinuities [8, 9]. However, these fringe projection methods apply to diffuse

surfaces only.

As for those specular surfaces, interferometry is a typical non-contact testing method with a

precision of down to sub-nanometers [10]. However, it is sensitive to environmental disturbances

and inconvenient for measuring complex surfaces due to its low dynamic range [11]. The phase

measuring deflectometry (PMD) method is a powerful optical technique for measuring complex

specular surfaces [12]. The fringe patterns displayed on a screen are reflected by the surface under
test and captured by a camera, and then the pixel correspondence between the screen and camera is

obtained according to their gray levels of phases. The normal vectors of the surface under test are

specified based on the law of reflection, and finally, the surface form is reconstructed from surface

gradients. The PMD possesses remarkable advantages such as high precision, high efficiency, high

dynamic range, and high robustness against environmental disturbances [13].

All the aforementioned methods need to retrieve phases from sinusoidal fringes, but 2 π

discontinuity occurs due to the periodicity of the sinusoidal and cosine functions. Then phase

unwrapping is required to obtain the true phase maps. There are primarily two kinds of methods,

namely the spatial phase unwrapping and temporal phase unwrapping. Spatial phase unwrapping

compares the phase values of adjacent pixels and eliminates the 2π discontinuities, such as the

diamond type algorithm [14], the quality-guided method [15], and the deep-learning-based method

[16]. It is efficient, but it unrobust to noise, leading to cumulative phase errors propagating from

high-noise regions to low-noise regions. Temporal phase unwrapping involves

projecting/displaying a series of fringe patterns with different frequencies. The phase associated

each pixel can be independently specified. Therefore it applies for discontinuous surfaces and the

phase errors will not propagate [17]. The most widely employed temporal methods is the three-

frequency heterodyne method [18,19].

Unfortunately, when measuring transparent components by deflectometry, the fringe images

reflected from and front and rear surfaces are superposed, making it difficult to establish pixel

correspondences. In addition, the low reflectivity of transparent elements results in a very low

signal-to-noise ratio (SNR) in the captured images, consequently reducing the measurement

accuracy. Existing methods for measuring transparent elements require a lot of fringe images to be

captured, or introducing additional hardware into the PMD setups [20, 21]. As a result, there is an

urgent demand for an efficient and convenient method for the deflectometric measurement of

transparent elements. Due to the excellent representativity of curve features, the curvelet transform

is adopted for fringe separation. The proposed method can significantly reduce the number of

images to be captured, avoid the demand for additional equipment, and achieve high-precision

measurement. The present work has two main contributions,

1) A fringe separation method is proposed for the deflectometric measurement of transparent

 elements based on the empirical curvelet transform. Fringe separation and correction are

realized simultaneously.

2) A weighted permutation entropy-based fusing method is proposed to recombine over-separated

modes and automatically allocate the fringes associated with the two surfaces of transparent

elements.

This paper is organized as follows. A brief introduction is presented in Section 1. Section 2

delineates the related work. The proposed method is introduced in Section 3 and numerically

investigated and compared in Section 4. Then it is experimentally demonstrated in Section 5, and

finally, the paper is concluded in Section 6.

2. Related work

A lot of work has been developed for the deflectometric measurement of transparent elements.

The reflection from the rear surface can be suppressed by using ultraviolet light as most mineral

glass and plexiglass are opaque to the radiation below the wavelength of 330 nm [22]. However,

building a ‘UV fringe projector’ would pose serious technical difficulties and high costs. In addition,

this method eliminates the backward-reflected signal, hence it could not measure double surfaces at

the same time.

In [23], the reflected light from both surfaces can be separated by displaying and scanning dots

or lines, but line scanning is often time-consuming and some areas cannot be detected. In [24], Hao

et al used binocular stereovision to measure the normal directions of transparent objects. The

consistency constraint of stereo normal vectors can be used to calculate the form of the transparent

element under test. However, due to the limitation of the binocular method, its measurement

accuracy is low, and additional uncertainty is introduced. In [25], a binary pattern method was

proposed to measure the front surface of a transparent element. It requires setting an appropriate

threshold to distinguish the regions associated with the rear surface, which is often not

straightforward to set. In [26], Jeong et al proposed a digital filtering method, in which the optical

layout and lighting source area are modified to bypass the generation of unwanted light. This method

is only applicable when there exists a large difference between the radii of curvature of the two

surfaces. In [27], the deflectometric measurement of transparent elements based on fringe separation

was proposed. The two-dimensional fast iterative filtering is developed to optimize the fringe
separation results. However, this method cannot separate the fringes with similar frequencies but

different directions, and the precision of fringe separation is not high, thereby resulting in mode

mixing. The combination of modes needs to be implemented manually.

The multi-frequency method is widely used for fringe separation, in which the phase is solved

iteratively by nonlinear optimization [28]. However, a disadvantage of this method is that plenty of

images are needed and appropriate initial values need to be provided in advance, and phase

discontinuity may occur due to the numerical singularity problem. Tao et al. [29] proposed an

envelope curve algorithm to eliminate parasitic reflection and obtain the phases associated with the

front and rear surfaces at the same time, but this algorithm needs to display hundreds of fringe

patterns with different frequencies. In [30], a method is proposed to obtain the initial conditions and

reduce the required number of images by about 8 times. However, this method is time-consuming

because the numerical optimization is implemented pixel by pixel, and it behaves robustly only for

relatively planar elements. Wang et al. [31, 32] proposed an algorithm to analyze harmonic

components in the power spectrum to unwrap the double-surface reflection, but this method applies

only to planar elements.

3. Methodology

3.1 Reflective phase measuring deflectometry for transparent elements

Unlike the scenario of measuring ordinary mirrors, the setup for measuring transparent elements

is shown in Fig. 1 (a). The propagation of light in the reflective phase measuring deflectometry

system can be divided into two paths. Starting from a point S1 on the screen, the light of the first

path is directly reflected via a point M1 on the front surface and captured by the camera; The light

of the second path, which starts from a point S2 on the screen, passes by a point M2 and is reflected

by a point N1 on the rear surface, refracted through M1, and then captured by the camera. The

resulting two sets of fringes are superposed with each other and the underlying ground-truth phase

cannot be obtained. When the light is approximately vertically incident, the reflectivity of glass is

about 4%. Therefore, the light intensity reflected by the rear surface at one time is slightly lower

than that by the front surface. On the contrary, the secondary reflection of light is much weaker,

thus it can be ignored. Fig. 1 (b) is a typical superposed fringe pattern caused by parasitic reflections.
 The captured image can be expressed as

I c = A + B1 cos j1 + B2 cos j 2 + h （1）

where Ic represents the superposed intensity, A is the background intensity, η stands for random

noise, B1 and B2 are the modulation coefficients of the reflected fringes from the front and rear

surfaces, and φ1 and φ2 are the corresponding phase values, respectively.

Fig. 1. PMD measurement of a transparent element. (a) PMD system, (b) superposed fringe pattern

When the fringes are correctly separated, phase retrieval and phase unwrapping can be conducted

to obtain absolute phase values and thus the correspondence relationship between camera pixels and

screen pixels can be established accordingly. The multi-step phase-shifting technique can be used

for phase retrieval [33]. Then, the absolute phases are determined using marker-assisted spatial

phase unwrapping and multi-frequency temporal phase unwrapping [34].

After the correspondence relationship between the camera pixels and screen pixels is established,

the Software Configurable Optical Test System (SCOTS) method is adopted to iteratively calculate

the X-direction gradient gx and Y-direction gradient gy at a measurement point M [35]. The surface

form could be reconstructed from the gradients [36]. In order to address the refraction of rays in the

transparent element, the form of the rear surface can be reconstructed reliably by using a hybrid

reflection-refraction model [28].

3.2 Empirical curvelet transform

The adaptive empirical wavelet transform proposed by Gillies et al [23, 24] defines a set of

bandpass filters in the frequency domain. It can automatically identify the Fourier support that best

separates the harmonic modes present in the analyzed signal. Their selections of support in the

Frequency domain are not specified as dyadic tiling but adaptively selected.

In a one-dimensional (1D) empirical wavelet transform (EWT), assuming that the signal is
composed of N single components. N wavelet filters, including a low-pass filter and N-1 bandpass

filters, are defined in the 1D Fourier domain based on ‘carefully selected’ supports. In this way, the

spectrum is divided into N continuous segments, with the n-th segment represented as Λn = [ωn-1,

ωn], n=1, 2, …, N, and ÈLn = [0, p ] , as shown in Fig. 2. t represents the width of the transition

band of filters. Several boundary partitioning methods can be adopted [23], e.g., the local maximum

of the spectrum amplitude is identified, and then, filters can be constructed in each frequency band

to extract those components of interest.

Fig. 2. Fourier spectrum decomposition and empirical wavelet basis construction.

The bidimensional Empirical Wavelet Transform (BEWT) is a two-dimensional (2D) extension

of the Empirical Wavelet Transform, which includes various extension methods, such as the Tensor,

Littlewood-Paley, and curvelets [37]. The empirical curvelet transform (ECT) is an appropriate fit

of curves, because the curvelet transform can provide a stable, efficient, and near-optimal

representation of discontinuities along smooth curves, and the twisted 2D sinusoidal fringes can be

viewed as consisting of an infinite number of smooth curves. The ECT first implements the pseudo-

polar Fourier transform to 2D data, and then divides the Fourier spectrum to obtain a set of scale

boundaries Ωω = {ωn}, n = 0, 1, × × , NS and a set of angular boundaries Ωθ = {θm}, m = 0, 1, …, Nθ.

When dividing the frequency domain, one can to detect the angles first and then find a set of scale

radii for each angular sector, as shown in Fig. 3(a), or detect the scale radii first and then detect the

angular sectors for each scale, as shown in Fig. 3(b). Of course, they can also be detected

independently.

Fig. 3. Different ways of curvelet tiling in the frequency domain.

The bandpass curvelet filters Ψn,m are polar wedges defined by the product of the radial windows
Wn and the polar windows Vm. Wn are defined as:

If n ≠ Ns − 1,

ì1 (1 + g )w n £ w £ (1 - g )w n +1 ,
ï
ï ép æ 1 n +1 ö ù
ïcos ê 2 b ç 2gw n +1 ( w - (1 - g )w ) ÷ ú (1 - g )w n +1 £ w £ (1 + g )w n +1 ,
ï ë è øû (2)
Wn (w ) = í
ï é p æ 1 n ö ù
ïsin ê 2 b ç 2gw n ( w - (1 - g )w ) ÷ ú (1 - g )w n £ w £ (1 + g )w n ,
ï ë è øû
ï
î0 otherwise

and if n = Ns − 1,

ì1 (1 + g )wNs -1 £ w ,
ï
ï ép æ 1 öù
WNs -1 (w) = ísin ê b ç ( w - (1 - g )wNs -1 ) ÷ú (1 - g )wNs -1 £ w £ (1 + g )wNs -1, (3)
ï ë 2 è 2gw
Ns -1
øû
ï0 otherwise
î

The polar window Vm are defined as:

ì1 q m + Dq £ q £ q m +1 - Dq ,
ï
ïcos é p b æ 1 (q - q m +1 + Dq ) ö ù q m +1 - Dq £ q £ q m +1 + Dq ,
ïï êë 2 çè 2Dq ÷ú
øû
Vm (q ) = í (4)
ïsin é p b æ 1 (q - q m + Dq ) ö ù q - Dq £ q £ q + Dq ,
m m
ï ê 2 çè 2Dq ÷ú
øû
ï ë
îï0 otherwise

The low-pass filter ϕ1 is defined as:

ì1 w £ (1 - g )w1 ,
ï ép æ 1 öù
ï
f1 (w ) = ícos ê b ç ( w - (1 - g )w1 ) ÷ ú (1 - g )w1 £ w £ (1 + g )w1 , (5)
ï ë 2 è 2gw
1
øû
ïî0 otherwise

where |·| denotes taking the absolute value. ω and θ are the polar coordinates in the Frequency

domain. β is an arbitrary function, with a suitable choice β(x) = x4(35 − 84x + 70x2 − 20x3). The

parameter γ is properly set to ensure that only two consecutive filters can overlap. The corresponding

sets of filter banks are given by

Fe c = {f1 ,{y n,m }n=1,2,×××, Ns -1;m=1,2,×××, Nq } (6)

where ϕ1 is the low-pass filter described in Eq. (5), and {Ψn,m} are the polar wedges. The curly
brackets {} denote a set. Finally, conduct filtering on the Fourier transform F of the input image

using the filter bank Фεc,

Wec (n, m) =áF, Fec ñ (7)

where Wεc represents the curvelet coefficients in the Fourier domain, and the angle brackets < >

denote the Hadamard product, i.e., element-wise multiplication. Then, the spatial filtering results can

be obtained by the inverse Fourier transform of Wεc.

3.3 Empirical curvelet transform for deflectometry

Ideal 2D sinusoidal fringes appear as a pair of conjugate symmetric points in the frequency

domain. However, due to various factors such as the curvedness of the fringes, frequency changes,

and spectral leakage, the fringe signals will disperse in the frequency domain and be accompanied

by false signals, as illustrated in Fig. 4.

Fig. 4. Examples of different fringe images and their Fourier spectra. (a) fringe images, (b) Fourier spectra, (c)

Fourier spectra after windowing.

The aperiodic truncation of the image will lead to remarkable spectral leakage, which in turn

seriously affects the detecting accuracy of the radial boundaries. As shown in Fig. 4, there are plenty

of horizontal and vertical artifact lines in the Fourier domain. Therefore, it is recommended to
specify the angular boundaries before the specifying radial boundaries. The pronounced spectral

leakage is suppressed by windowing. The Hamming window, with its first sidelobe attenuation of -

42 dB and a sidelobe attenuation rate of -20 dB/decade, is adopted for such a purpose. It is defined

as

ì 2p x
ï0.54 - 0.46cos 0 £ x £ N -1
Hamming(x) = í N -1 (8)
îï 0 else

where x represents the lateral coordinates. Multiplying the 2D window function by the original

image f as

f w = f e (Hamming(x) ´ Hamming(y )) (9)

where ⊙ represents the Hadamard product, and y is the vertical coordinate. Then perform the

pseudo-polar Fourier transform on fw. For each angle, calculate the average of those points with the

top 5% highest magnitudes as an indicator for spectrum segmentation. This approach ensures that

even if the sidelobe attenuation rate is low, the spectrum segmentation will not be significantly

affected.

As described in [38], a fine-to-coarse algorithm is adopted to detect a set of angular boundaries.

This is a non-parametric histogram segmentation method that does not require any prior assumptions

of modes in the histogram. An adequacy test called 'meaningful rejections' is proposed to determine

whether two consecutive separated modes belong to the same mode. After that, a set of angular

boundaries Ωθ = {θm}, m =1, 2, × × × , Nθ, is obtained, where Nθ is the number of the angular segments.

Furthermore, if the angular interval Ωθ contains the fringe component of interest, the spectral

intensity within this interval should be one of the first two maximum values. Therefore, in order to

remove the superfluous boundaries, the first two maximum values of Nθ are detected, and then the

left and right borders Ωθ' closest to them are preserved, while the others are deleted. Within each

angular interval, a fine-to-coarse algorithm is adopted again to detect radial supports to obtain a set

of scale boundaries Ωω.

It is worth noting that windowing is only applied to enhance the detection accuracy of the Fourier

supports, rather than participating in filtering directly. In practice, the original image f is filtered in
the frequency domain using the wedge filters Ψn,m formed by Ωω and Ωθ', as presented in Eq. (6).

3.4 Recombine the modes of excessive separation

The results of separating an image with different filters are called modes. In some cases, excessive

separation of modes is inevitable, which leads to fringe components missing, as shown in Fig. 5.

Fig. 5. Examples of excessive separation of modes

They need to be recombined together in the spatial domain to recover the complete fringe patterns.

A two-stage method is proposed to remove invalid modes, reassemble identical modes, and

automatically distinguish the front and rear fringe patterns from multiple modes.

Before recombining the patterns, obviously meaningless modes should be removed first. The

Weighted Permutation Entropy (WPE) is a nonlinear statistical index used to estimate the

complexity of time series [39]. The more regular the time series is, the smaller the corresponding

permutation entropy is. Since sinusoidal fringes still appear as sinusoidal when projected

horizontally and vertically, WPE can be simply generalized to a 2D form to represent the degree of

disorder and abrupt changes of 2D fringe patterns. The separated fringe patterns can be regarded as

a set of row or column signals. The target fringe patterns are approximately sinusoidal with a

uniform amplitude and low entropy, while noise is chaotic with high entropy. The variance of each

pattern can be calculated after excluding noise. The reason for using a variance instead of directly

calculating energy is that low-frequency backgrounds may also have relatively high energy levels.

The 1D WPE algorithm consists of the following steps:

Step 1: For a 1D sequence {xn}, n = 1, 2, × × × , N, where N is the data length, map the signal to a

m-dimensional space to obtain K = N−(m−1)λ subsequences as:

ì X 1 = x1 , x1+ l , × × ×, x1+ ( m -1) l

ï
ï M
í X j = x j , x j + l , × × ×, x j + ( m -1) l
ï M (10)
ï
î X K = x K , x K + l , × × ×, xK + ( m -1) l

where j=1, 2, …, K. m is the embedded dimension, and λ is a time delay.

Step 2: Elements in each row can be arranged in an ascending order as

 Aj = ëé x j +( j1 -1)l , x j +( j2 -1)l , ×××, x j +( jm -1)l ûù (11)

where j1, j2, ..., jm represent the column indices of a subsequence. If the columns are of an equal

length, the elements are listed according to their indices. Then, each subsequence Aj can be mapped

to an ascending symbol permutation with a size relationship as

S j = { j1 , j2 , ×××, jm } (12)

Obviously, m elements have at most m! different permutations, and the set of all the symbol

permutations is denoted as Π.

Step 3: All permutations are counted and the probability of a certain permutation Sj is calculated as

ål
j =1
Su ÎSv ( X j )w j
pw ( wSv ) = K (13)
å lSu ÎP ( X j )w j
j =1

where

lsu Îs ( X j ) = 1,0, Su Î S
Su Ï S (14)

where Sv is a symbol permutation type in Π, v = 1, 2, …, m!. Su is the symbol permutation type of

Xj, and S is a particular symbol permutation or a set of symbol permutations. wi is the weight of the

reconstruction component xj.

2
1 m 1 m
wj = å j +( k -1)l m å
m k =1
[ x -
k =1
x j + ( k -1) l ] (15)

Step 4: WPE is then computed as

m!
WPE  xn , m, l  = -å pw ( wSi ) log( pw ( wSi )) (16)
i =1

For an image, we sample T 1D sequences {xn,t}, n = 1, 2, × × × , N; t = 1, 2, × × × , T, in rows and

columns where N is the data length. The WPE is calculated for each sequence separately, and then

their average is calculated to extend it into two dimensions.

1 T
WPE2D(m, l ) = å WPE( xn,t , m, l )
T t =1
(17)

A mode with its WPE2D value greater than 0.6 can be regarded as noise and removed. Then,

calculate the variance of the remaining modes as

 1 N
Var = å (x i - m )2
M - 1 i =1
(18)

where xi represents pixel values, M is the total number of pixels, and μ is the arithmetic average of

{xi}.

The two modes whose variance rank the first and the second are chosen as the initial fringe

patterns of the front and rear surfaces, respectively denoted as F1 and F2. The remaining modes are

denoted as F3, F4, …, and FL, where L is the total number of separated modes. When WPE2D is

minimized, the fringe pattern undergoes no mutations, and the level of disorder is at its minimum,

indicating the complete integration of fringe information. The modes recombination algorithm is

shown in algorithm 1.

Algorithm 1. Mode recombination algorithm

Input: F1, F2, …, FL.
Output: F1', F2'.
1: for i=1 to 2 do
2: Fi'=Fi
3: for j = 3 to L do
4: Compute WPE2D(Fi') and WPE2D(Fi'+Fj)
5: if WPE2D(Fi'+Fj) < WPE2D(Fi') then
6: Fi' = Fi+Fj
7: end if
8: end for
9: end for

The flowchart diagram is shown in Fig. 6.

4. Numerical investigation and comparison

In this section, some numerical experiments are presented first. The results of the proposed

method are compared with those associated with the empirical curvelet transform and those of

mainstream mode decomposition algorithms, respectively.

 Fig. 6. Flow-chart diagram of the entire algorithm

4.1 The proposed method vs. ECT

Fig. 7. Superposed images and their boundaries in the Frequency domain. (a)an example image, (c) boundaries of
(a) detected by ECT, (d) enlarged partial view of (c), (e) the boundaries of (a) detected by our method, (f) enlarged
partial view of (e). The right part shows another example

A set of 200 superposed fringes is generated according to Eq. (1), including fringe patterns, noise,

and background. Among them, randomly generated three-ordered polynomials are used to twist the

fringes by adjusting the fringe width distribution. The fringe period ranges from 20 to 128 pixels,

and the background consists of a two-dimensional Gaussian distribution with a variance in a range

250~550. The noise is a combination of normally distributed noise and salt-pepper impulsive noise,
and the peak signal-to-noise ratio (PSNR) is between 10 and 25 dB. The settings of ECT described

in Subsection 2.2 are applied to decompose and de-noise the fringe patterns. The ECT method does

not give an answer on how to recombine the possible separated modes, and this task needs to be

implemented manually. Two superposed fringe patterns are selected, as shown in Fig. 7(a), and (b).

Fig. 7(c), (g), (e) and (i) show the boundaries detected using EWT and the proposed method. It is

obvious that the direct use of the ECT algorithm will produce plenty of unnecessary patterns, and it

will be difficult to identify and recombine these massive patterns.

Fig. 8. Histogram of SSIM value of 200 images. (a) SSIM value of all pixels of 200 images. (b) SSIM value of 200
images after 30 pixels near the edges of the image are deleted.

On the other hand, the proposed method behaves superiorly on the aspect of accuracy of mode

separation, and it is more inclined to fine segmentation in the regions containing the signal, as shown

in Fig. 7 (f) and (j). In most cases, the proposed algorithm can separate the fringe patterns completely,

with little information leaked into the background. The structured similarity index measure (SSIM)

is used to quantitatively measure the image quality, as shown in Fig. 8 (a). A higher SSIM value

indicates more complete separation. It can be found that the mean value of the SSIMs obtained by

the proposed algorithm is about 0.84, and if the edge effect is not considered, this value will be

further increased as shown in Fig. 8 (b). In terms of computation time, this algorithm takes only 10

seconds to compute a single channel image of 512 × 512 pixels in MATLAB on a laptop with AMD

Ryzen 7 5800H CPU.

4.2 The proposed method vs. well-established mode decomposition algorithms

The proposed approach is compared with mainstream modal decomposition methods, namely

2D-TV-VMD [30], fpFIF2 [41], and BEMD [42], and the results are shown in Fig. 9. The SSIM

values of different results are presented in Table 1.

The 2D-TV-VMD algorithm has 13 parameters, which need to be finely tuned to achieve
satisfactory results, and the number of modes needs to be specified a priori. The 2D-TV-VMD has

the lowest level of automation but possesses high robustness. fpFIF2 separates fringe patterns in the

same direction and different frequencies effectively, but inaccurate parameter setting will reduce

the accuracy of separation, which means that the parameters need to be adjusted according to the

actual situation. Since fpFIF2 employs the Fokker-Planck filter, which lacks directional analysis

capabilities, it performs poorly in separating superposed fringes with similar frequencies but in

different directions, as shown in Fig. 9 (c), (d). BEMD does not require any parameter adjustments

and it has the highest level of automation but produces poor and unstable separation results. As it

does not incorporate any prior information about fringes, the AM-FM components separated by

BEMD may not be as expected. Interestingly, our method consistently achieves good fringe

separation results in various situations, with clear fringe patterns and minimal mode mixing.

Another advantage is that it can automatically recombine possibly over-separated modes, allowing

end-to-end output of results. During experimentation, this method performs well without manually

adjusting any parameters.

Fig. 9. Comparison of fringe separation quality with well-established mode-decomposing algorithms. (a), (b), (c),
(d) represent different examples of simulated superposed fringe patterns

The fringe patterns separation is a step in deflectometry for transparent elements. The ultimate

measure of the separation results is the accuracy of the retrieved phases associated with the double

surfaces. As a decrease in contrast due to the loss of energy may not necessarily result in significant

errors in the phase, whereas perturbations in the shape of the fringes after filtering may lead to errors

in the pixel correspondences.

To further validate the performance of the proposed algorithm, ray tracing is conducted to obtain

the fringe patterns associated with different plano-convex lenses, plano-concave lenses, biconvex

lenses, biconcave lenses, and crescent lenses. The four-step phase-shifting technique is applied for

phase retrieval and then the Mean Squared Error (MSE) and Max Error (ME) between the separation

results and the ground truth are adopted for quantitative comparison. The calculation methods for

MSE and ME are shown in Eq. (20) and Eq. (21), respectively.

1 p
MSE = å ( yi - Yi )2
p i =1
(20)

ME = max | y i - Yi | (21)
i

where p is the total number of pixels, and yi is the retrieved phase of the i-th pixel, and Yi is the
 associated ground truth. In the fringe patterns, background and random noise are introduced as

described in Section 3.1. The separation results of different methods are listed in Tables 2 to 6.

Table 1. SSIMs of different fringe separation methods

Fringes 2D-TV-VMD FpFIF2 BEMD Ours
0.8992 0.8279 0.5978 0.8279
(a)
0.7230 0.7753 0.5249 0.8053
0.7677 0.8546 0.5825 0.8273
(b)
0.7919 0.8443 0.6052 0.8720
0.8006 0.5784 0.2558 0.8271
(c)
0.5154 0.4523 0.3621 0.8126
0.7360 0.4378 0.3176 0.8457
(d)
0.6457 0.2160 0.2665 0.8244

Table 2. The phase quality of the separated fringes of crescent spherical lenses
Radius of 2D-TV-VMD fpFIF2 BEMD Ours
Thickness
Surface curvature MSE ME MSE ME MSE ME MSE ME
(mm)
(mm) (rad) (rad) (rad) (rad) (rad) (rad) (rad) (rad)
front 200 0.030 0.117 0.014 0.083 0.772 4.999 0.036 0.168
5
rear 500 0.072 0.213 0.085 0.623 0.628 4.999 0.029 0.133
front 500 0.823 4.99 0.158 0.489 1.856 4.998 0.078 0.378
7
rear 300 0.508 1.021 0.274 0.766 1.381 4.999 0.050 0.198
front 600 0.035 0.161 0.017 0.087 0.942 5.000 0.031 0.159
4
rear 200 0.489 4.15 0.037 0.123 1.561 4.999 0.068 0.345

Table 3. The phase quality of the separated fringes of plano-concave spherical lenses

Radius of 2D-TV-
Thickness fpFIF2 BEMD Ours
Surface curvature VMD
(mm) MSE ME MSE ME MSE ME MSE ME
(mm)
(rad) (rad) (rad) (rad) (rad) (rad) (rad) (rad)
front 300 0.011 0.039 0.016 0.134 0.379 4.988 0.026 0.154
6
rear 20000 0.145 0.338 0.024 0.107 0.165 2.804 0.031 0.134
front 400 0.376 0.874 0.014 0.065 0.445 4.997 0.023 0.121
3
rear 20000 0.228 0.616 0.039 0.134 0.398 4.986 0.021 0.152
front 250 0.117 0.310 0.011 0.073 0.774 4.449 0.025 0.116
10
rear 20000 0.324 0.986 0.014 0.063 0.336 4.986 0.026 0.096

Table 4. The phase quality of the separated fringes of plano-convex spherical lenses
Radius of 2D-TV-VMD fpFIF2 BEMD Ours
Thickness
Surface curvature MSE ME MSE ME MSE ME MSE ME
(mm)
(mm) (rad) (rad) (rad) (rad) (rad) (rad) (rad) (rad)
front 300 0.030 0.117 0.017 0.076 0.178 2.384 0.026 0.104
3
rear 20000 0.028 0.063 0.043 0.178 1.103 4.999 0.033 0.176
front 400 0.355 0.665 0.029 0.096 0.626 4.997 0.033 0.210
7
rear 20000 0.021 0.064 0.042 0.186 1.021 4.994 0.040 0.153
front 250 0.170 0.463 0.023 0.071 0.191 4.998 0.044 0.252
10
rear 20000 0.019 0.062 0.028 0.084 0.202 4.986 0.034 0.154
 Table 5. The phase quality of the separated fringes of biconvex spherical lenses
Radius of 2D-TV-VMD fpFIF2 BEMD Ours
Thickness
Surface curvature MSE ME MSE ME MSE ME MSE ME
(mm)
(mm) (rad) (rad) (rad) (rad) (rad) (rad) (rad) (rad)
Front 300 0.069 0.241 0.018 0.099 0.315 4.489 0.022 0.123
9
Rear 500 0.134 0.409 0.018 0.072 0.269 4.998 0.012 0.063
Front 400 0.070 0.222 0.021 0.098 0.175 1.313 0.043 0.159
11
Rear 300 0.542 5.000 0.024 0.200 0.071 0.853 0.047 0.200
Front 300 0.079 0.248 0.051 0.229 0.126 0.975 0.035 0.144
5
Rear 600 0.161 0.449 0.038 0.158 0.217 4.763 0.032 0.169

Table 6. The phase quality of the separated fringes of biconcave spherical lenses
Radius of 2D-TV-VMD fpFIF2 BEMD Ours
Thickness
Surface curvature MSE ME MSE ME MSE ME MSE ME
(mm)
(mm) (rad) (rad) (rad) (rad) (rad) (rad) (rad) (rad)
Front 300 0.033 0.080 0.042 0.167 0.824 0.813 0.028 0.162
15
Rear 300 0.025 0.065 0.048 0.379 0.202 4.089 0.024 0.122
Front 400 0.028 0.060 0.026 0.089 0.774 5.000 0.037 0.229
10
Rear 300 0.032 0.996 0.301 0.160 0.440 4.994 0.025 0.153
Front 250 0.025 0.093 0.013 0.056 0.802 4.999 0.019 0.097
8
Rear 600 0.024 0.104 0.024 0.119 0.256 4.972 0.065 0.363
The above experiments prove the effectiveness and superiority of the proposed method, as well

as the decomposition accuracy. The MSE results are comparable to those of fpFIF2, with an average

phase MSE of less than 0.08 rad.

5. Experimental demonstration and discussions

A deflectometry experimental system is set up. The system consists of a Basler acA1920-150µm

USB 3.0 camera with a resolution of 1200×1920 pixels and a pixel size of 4.8 µm, an Apple iPad

Mini2 screen with a resolution of 2048×1536 pixels, as illustrated in Fig. 10. Two transparent

elements are used for measurement, one of which is a plano-convex spherical lens with a convex

curvature of radius of 361.76 mm, an aperture of 76.2 mm, and a central thickness of 7 mm, and the

other is a plano-convex cylindrical lens with a convex radius of curvature of 129.15 mm, a central

thickness of 4 mm and a size of 40×40mm. The refractive indices of both elements are measured

with an Abbe refractometer, which turns out to be 1.5192. The system is calibrated using an

integrated calibration board, and the components under test are fixed at the same position as the

calibration board.
 Fig. 10. The measurement system.

In order to mitigate the influence of random noise and fringe separation error, a four-step phase-

shifting method combined with three-frequency heterodyne is adopted to obtain the wrapped phase

and perform temporal phase-unwrapping [20]. A total of 24 orthogonal fringe patterns of three

different frequencies are displayed, and then the images captured by the camera are cropped to

500×500 pixels for the plano-convex spherical lens and 300×300 pixels for the plano-convex

cylindrical lens to speed up the calculation and avoid the influence of the background. The front

surface of the element is reconstructed using the SCOTS method and Zernike polynomials. It is

worth noting that every 4 of the 24 fringe patterns have the same frequency but different phases,

thus the spectrum boundaries need to be detected only 6 times, and then the fringe patterns with

different phases of the same frequency are separated using the same filter bank. It takes less than 8

seconds to separate a superposed fringe pattern, and the entire experiment takes about 3 minutes.

5.1 Fringe separation results

Fig. 11 shows the fringe patterns before and after fringe separation associated with a spherical

lens. Because the shape of the fringes is simple and the frequency difference is large, the fringes are

effectively separated and there is almost no mode mixing, except for the method BEMD. The

empirical mode decomposition method based on the mathematical envelope cannot accurately

separate the required modes.

Fig. 11. Experiment results of the plano-convex spherical lens. (a) superposed image in the x direction. (b) s
superposed image in the y direction

Fig. 12. Experiment results of the plano-convex cylindrical lens. (a) superposed image in the x direction. (b)
superposed image in the y direction
 Fig. 12 shows the fringe patterns associated with a cylindrical lens before and after separation

using different methods. The fringe patterns of the front and rear surfaces have similar frequencies,

but obviously different directions, as shown in Fig. 12(b). The VMD method requires adjusting a

large number of parameters, and it is difficult to establish a mapping between fringes and

corresponding parameters, which is inconvenient to apply. Fig. 13 illustrates all the modes separated

by fpFIF2 in Fig. 12(b). BIMF6 suffers from severe mode mixing, leading to a significant

degradation of fringes. On the contrary, the proposed method still exhibits robustness.

Fig. 13. All modes separated by fpFIF2 in Fig .12 (b).

5.2 Comparative experiment results

The multi-frequency fringe separation approach [28] the neural network-assisted fast iterative

filtering method [27] are adopted for comparison on the transparent elements described in

Subsection 5.1. When using the multi-frequency method, it is necessary to perform reverse ray

tracing to obtain the initial phase values, which is quite troublesome and time-consuming. Moreover,

it requires capturing 32 fringe images. As for the latter, the 2D fast iterative filtering is used for

coarse fringe separation, and a neural network is established for refinement, as depicted in Fig. 14.

By comparing the separation results of fpFIF2 described in Fig. 11 and Fig. 12, it is evident that

after neural network optimization, the adverse effects caused by noise, background interference,

modulation loss and mode mixing are significantly reduced. Due to the poor fringe separation results
of fpFIF2, the fringes of the cylindrical lens still have some fluctuations caused by noise and mode

mixing after neural network optimization. Therefore, the accuracy of surface form for the cylindrical

lens remains relatively low. Additionally, the mode decomposition results of the referenced method

require manual combination, which becomes problematic when there exist numerous modes or over-

decomposition.

Fig. 14. Comparative fringe separation results of the spherical lens and cylindrical lens. (a) fringes of the front
surface of the spherical lens, (b) fringes of the rear surface of the spherical lens, (c) fringes of the front surface of
the cylindrical lens, (d) fringes of the rear surface of the cylindrical lens
The phase unwrapping results of the spherical lens are shown in Fig. 15 (a), (c), and (e), while

those of the cylindrical lens are shown in Fig. 15 (b), (d), and (f). The phase maps obtained with

[28] exhibit phase jumps and data missing, as shown in Fig. 15 (c) and (d). After phase retrieval,

the forms of the front surfaces of the two lenses are reconstructed and compared with those of the

LuphScan profilometer. The measured forms and relative deviations of the spherical lens are

depicted in Fig. 16, whereas those of the cylindrical lens are depicted in Fig. 17.

The proposed method achieves a root mean square error (RMSE) of 0.073 μm and a peak-to-

valley error (PVE) of 0.436 μm for the spherical lens, and 0.322 μm and 1.471 μm for the cylindrical

lens, which demonstrate the superiority of the proposed method, as listed in Table 7.

Table 7. Experimental results of the plano-convex spherical lens and the plano-convex cylindrical lens
Methods Lens RMSE/μm PVE/μm
The Spherical 0.073 0.436
proposed Cylindrical 0.322 1.471
Spherical 0.288 1.81
[28]
Cylindrical 0.774 4.23
Spherical 0.119 0.438
[27]
Cylindrical 1.01 7.81
 Fig. 15. Experiment results of the front surface of plano-convex cylindrical lens. (a) the phase of spherical lens
obtained by the proposed method, (b) phase of cylindrical lens obtained by the proposed method, (c) phase of
spherical lens obtained by [28], (d) phase of cylindrical lens obtained by [28], (e) phase of spherical lens obtained
by [27], (f) phase of cylindrical lens obtained by [27].

It is worth noticing that some high-frequency errors occur, as depicted in Fig. 15 (b) and (f). This

could be because some high-frequency components are falsely allocated into the noise, leading to a

mismatch between the camera pixels and the screen pixels. As a result, the surface shape calculated

by SCOTS will be erroneous, giving rise to high-frequency periodic errors. In addition, the increase

in the measurement error of the cylindrical lens is mainly caused by inaccurate calibration and

systematic error, as the position of the cylindrical lens is difficult to specify precisely.

Fig. 16. Experiment results of the front surface of the plano-convex lens. (a) form of the proposed method, (b)
form error of (a), (c) form measured by [28], (d) form error of (c), (e) form measured by [27], (f) form error of (e)
 Fig. 17. Experiment results of the front surface of the plano-convex cylindrical lens. (a) form measured by the
proposed method, (b) form error of (a), (c) form measured by [28], (d) form error of (c), (e) form measured by
[27], (f) form error of (e)

5.3 Discussions

The proposed method not only exhibits higher accuracy but also avoids phase jumps and

omissions. More importantly, unlike the multi-frequency method, it neither requires complex ray

tracing to determine initial phase values nor necessitates capturing hundreds of images. Even

without using neural networks for optimization, higher accuracy can be achieved compared to [27],

while significantly outperforming it in measuring the cylindrical lens. Furthermore, the proposed

method operates under identical experimental conditions with those of traditional PMD techniques,

requiring only one or two cameras, as well as a screen and a fixture to hold the transparent element.

This avoids the demand for auxiliary devices described in [24] and [26].

Fig. 18. An example of failure. (a) the superposed patterns, (b)the frequency domain image of (a)

Looking back at Fig. 8, there are a few images with very low SSIM, as shown in Fig. 18 (a). The

filter bank is shown in Fig. 18 (b). Although the proposed method is fast and robust in most cases,

not in all cases the superposed fringe patterns can be well separated. If the elements are very thin or
flat, or the two surfaces are nearly parallel to each other, the fringe frequencies of the two surfaces

are almost identical, as shown in Fig. 18 (b). For two fringes with very close frequencies, due to

resolution limitations in the frequency domain, the proposed method cannot separate them properly,

which results in unavoidable mode mixing.

In addition, the quality of the separated fringe patterns is poor at the edge. This is a common

problem of convolution-based digital filtering methods. Sudden truncation of the image boundary

will produce ringing effects and pattern distortion in the spatial domain, which will become serious

in the case of low frequencies. Methods such as image expansion can alleviate the degree of spatial

distortion, but they cannot completely eliminate its influence [43].

We examined various image padding techniques, such as zero padding, replication padding, and

anti-reflective padding [44]. Fig. 19 (a) illustrates superposed fringes, with (b) and (c) representing

ground truth, and (d) and (e) displaying error maps for fringe separation without padding. Fig. 20,

Fig. 21 and 22 demonstrate error maps associated with zero padding, replication padding, and anti-

reflective padding, respectively. Table 8 presents the MSE errors with respect to the ground truth.

It is evident that proper image expansion techniques can marginally enhance fringe separation

quality. Replication padding and anti-reflective padding are more effective than zero padding.

However, excessive padding can lead to a decline in the quality of fringe separation.

Fig. 19. Boundary effect on separation results. (a) superposed fringe pattern, (b) ground truth of fringe 1, (c)
ground truth of fringe 2, (d) error map of fringe 1, (e) error map of fringe 2

Table 8. Effects of different padding methods and padding sizes on fringe separation results
 Methods Fringe 0 pixels 6 pixels 12 pixels 48 pixels
No 1 0.0109 - - -
Padding 2 0.0126 - - -
1 - 0.0105 0.0102 0.0109
Zero
2 - 0.0120 0.0119 0.0145
1 - 0.0099 0.0097 0.0112
Replicate
2 - 0.0109 0.0096 0.0127
Anti- 1 - 0.0121 0.0092 0.0135
reflective 2 - 0.0161 0.0117 0.0138

Fig.20. Fringe separation error of "zero padding". (a) fringe 1 with 6 pixels padding, (b) fringe 2 with 6 pixels
padding, (c) fringe 1 with 12 pixels padding, (d) fringe 2 with 12 pixels padding, (e) fringe 1 with 48 pixels
padding, (f) fringe 2 with 48 pixels padding

Fig. 21. Fringe separation error of "replicate padding". (a) fringe 1 with 6 pixels padding, (b) fringe 2 with 6 pixels
padding, (c) fringe 1 with 12 pixels padding, (d) fringe 2 with 12 pixels padding, (e) fringe 1 with 48 pixels
padding, (f) fringe 2 with 48 pixels padding

In the practical application, structural light profilometry and other methods involve fringe analysis
may suffer from issues of low fringe quality and fringe superposition. As a result, the proposed

method can be applied in these measuring methods to suppress disturbance and extract underlying

fringe components. For the convenience of the readers, the data and codes of the developed

algorithm are openly available at https://github.com/YangDDgood/Deflectometry.

Fig. 22. Fringe separation error of "anti-reflective padding". (a) fringe 1 with 6 pixels padding, (b) fringe 2 with 6
pixels padding, (c) fringe 1 with 12 pixels padding, (d) fringe 2 with 12 pixels padding, (e) fringe 1 with 48 pixels
padding, (f) fringe 2 with 48 pixels padding

6. Conclusions

In this paper, an empirical curvelet transform method is proposed to separate the superposed

fringes associated with transparent elements. It assumes that the two fringe patters can be well

separated in the frequency domain. The proposed method can obtain the fringe patterns of the two

surfaces at the same time, greatly reducing the time of measurement and data processing, and

realizing full automation without adjusting any system configurations. In addition, numerical

experiments and two measurement experiments are conducted to quantitatively demonstrate the

validity of this method. In the case of separating low SNR fringe patterns and fringe patterns in

different directions, it proves attractive robustness and separation accuracy. These advantages

suggest that this method is a valuable alternative to the current mainstream multi-frequency phase-

shifting methods and it is promising for automated measurement in advanced manufacturing.

Declarations

The authors declare no competing interests.

References

[1] G.M. Williams, H. Akhavan, C. Dupuy, P. Harmon, Additive manufacturing of freeform optics for
defense applications, 2021 IEEE Research and Applications of Photonics in Defense Conference, (2021)
1-2.
[2] C. Peixoto, P.T. Valentim, P.C. Sousa, D. Dias, C. Araújo, D. Pereira, C.F. Machado, A.J. Pontes, H.
Santos, S. Cruz, Injection molding of high-precision optical lenses: a review, Precision Engineering, 76
(2022) 29-51.
[3] R.J. Bensingh, M. Rajendra, S.R. Boopathy, C. Jebaraj, Injection molding process optimization of a
bi-aspheric lens using hybrid artificial neural networks (ANNs) and particle swarm optimization (PSO),
Measurement, 134 (2019) 359-374.
[4] H. Jing, C. King, D. Walker, Measurement of influence function using swing arm profilometer and
laser tracker, Optics Express, 18 (2010) 5271-5281.
[5] H. Guo, H. Zhou, P. Banerjee, Single-shot digital phase-shifting Moiré patterns for 3D topography,
Appl. Opt., 60 (2021) A84-A92
[6] X. Su, W. Chen, Fourier transform profilometry: a review, Optics and Lasers in Engineering, 35
(2001) 263-284.
[7] Y. Liu, Y. Fu, Y. Zhuan, K. Zhong, B. Guan, High dynamic range real-time 3D measurement based
on Fourier transform profilometry, Optics & Laser Technology, 138 (2021) 106833.
[8] M. Lu, X. Su, Y. Cao, Z. You, M. Zhong, Modulation measuring profilometry with cross grating
projection and single shot for dynamic 3D shape measurement, Optics and Lasers in Engineering, 87
(2016) 103-110.
[9] M. Zhong, F. Chen, C. Xiao, Y. Yang, Y. Wei, Noise reduction in modulation measurement
profilometry based on the wavelet transform method, 57 (2018) 054102.
[10] L. Zhang, R. Liu, J. Wu, Z. Cheng, S. Zhou, J. Li, B. Yu, Freeform surface adaptive interferometry
assisted with simulated annealing-hill climbing algorithm, Measurement, 181 (2021) 109597.
[11] X. Wu, L. Zhu, F. Fang, X. Zhang, Research on the quality control technology of micro-topography
machining based on in situ white light interferometry, Measurement, 220 (2023) 113257.
[12] M. W. Knauer, J. Kaminski, G. Hausler, Phase measuring deflectometry: a new approach to measure
specular free-form surfaces, Proc. SPIE, 5457 (2004) 366-376.
[13] Z. Zhang, C. Chang, X. Liu, Z. Li, Y. Shi, N. Gao, Z. Meng, Phase measuring deflectometry for
obtaining 3D shape of specular surface: a review of the state-of-the-art, Optical Engineering, 60 (2021)
020903.
[14] W. Wan, J. Su, L. Yang, J. Xu, A phase unwrapping algorithm for image processing of interferogram,
Journal of Applied Optics, 32 (2011) 70-74.
[15] M. Arevalillo-Herráez, F. R. Villatoro, M. A. Gdeisat, A robust and simple measure for quality-
guided 2D phase unwrapping algorithms, IEEE Transactions on Image Processing, 25 (2016) 2601-2609.
[16] S. Feng, C. Zuo, Y. Hu, Y. Li, Q. Chen, Deep-learning-based fringe-pattern analysis with uncertainty
estimation, Optica, 8 (2021) 1507-1510.
[17] X. He, K. Qian, A comparative study on temporal phase unwrapping methods in high-speed fringe
projection profilometry, Optics and Lasers in Engineering, 142 (2021) 106613.
[18] Z. Niu, X. Xu, X. Zhang, W. Wang, Y. Zhu, J. Ye, M. Xu, X. Jiang, Efficient phase retrieval of two-
directional phase-shifting fringe patterns using geometric constraints of deflectometry, Optics Express,
27 (2019) 8195-8207.
[19] B. Wang, R. Wang, J. Zhang, L. Wang, X. Zhang, Dual-frequency phase unwrapping method for
anti-jump errors of wrapped phase, Integrated Ferroelectrics, 226 (2022) 72-81.
[20] Y. Xu, F. Gao, X. Jiang, A brief review of the technological advancements of phase measuring
deflectometry, PhotoniX, 1 (2020) 14.
[21] X. Zhang, D. Li, R. Wang, Screen-monitored stitching deflectometry based on binocular stereo
vision, Measurement, 207 (2023) 112440.
[22] D. Sprenger, C. Faber, M.C. Seraphim, G. Häusler, UV-deflectometry: no parasitic reflections, Proc.
DGaO, 111 (2010) A19.
[23] T. Su, W.H. Park, R.E. Parks, P. Su, J.H. Burge, Scanning long-wave optical test system - a new
ground optical surface slope test system, Proc. SPIE, 8126 (2011) 100-109.
[24] Z. Hao, Y. Liu, Transparent object shape measurement based on deflectometry, Proceedings, 2
(2018) 548.
[25] R. Wang, D. Li, K. Xu, X. Zhang, P. Luo, Parasitic reflection elimination using binary pattern in
phase measuring deflectometry, Optics Communications, 451 (2019) 67-73.
[26] B. Jeong, S. Ordones, H. Quach, D. Kim, H. Choi, Digital filtering of ghost signal in phase
measuring deflectometry, Opt. Lett., 48 (2023) 1642-1645.
[27] T. Chen, P. Yang, X. Zhang, W. Lang, Y. Chen, M. Xu, Separation of fringe patterns in fast
deflectometric measurement of transparent optical elements based on neural network-assisted fast
iterative filtering method, Advances in Manufacturing. 2024, https://doi.org/10.1007/s40436-024-
00509-w
[28] J. Ye, Z. Niu, X. Zhang, W. Wang, M. Xu, Simultaneous measurement of double surfaces of
transparent lenses with phase measuring deflectometry, Optics and Lasers in Engineering, 137 (2021)
106356.
[29] S. Tao, H. Yue, H. Chen, T. Wang, J. Cai, Y. Wu, L. Yong, Elimination of parasitic reflections for
objects with high transparency in phase measuring deflectometry, Results in Physics, 15 (2019) 102734.
[30] Y. Leung, L. Cai, Untangling parasitic reflection in phase measuring deflectometry by multi-
frequency phase-shifting, Applied Optics, 61 (2022) 208-222.
[31] W. Zheng, D. Li, R. Wang, X. Zhang, R. Ge, L. Yu, W. Zhao, Parasitic reflection separation
deflectometry based on harmonic analysis, Measurement, 203 (2022) 111864.
[32] R. Wang, D. Li, L. Li, K. Xu, L. Tang, P. Chen, Q. Wang, Surface shape measurement of transparent
planar elements with phase measuring deflectometry, Optical Engineering, 57 (2018) 104104.
[33] P.S. Huang, S. Zhang, Fast three-step phase-shifting algorithm, Applied Optics, 45 (2006) 5086-
5091.
[34] L. Huang, M. Idir, C. Zuo, A. Asundi, Review of phase measuring deflectometry, Optics and Lasers
in Engineering, 107 (2018) 247-257.
[35] P. Su, R.E. Parks, L. Wang, R.P. Angel, J. H. Burge, Software configurable optical test system: a
computerized reverse Hartmann test, Appl. Opt., 49 (2010) 4404-4412.
[36] C. Schulze, D. Naidoo, D. Flamm, O.A. Schmidt, A. Forbes, M. Duparré, Wavefront reconstruction
by modal decomposition, Optics express, 20 (2012) 19714-19725.
[37] J. Gilles, me, G. Tran, S. Osher, 2D empirical transforms. wavelets, ridgelets, and curvelets revisited,
SIAM Journal on Imaging Sciences, 7 (2014) 157-186.
[38] J. Delon, A. Desolneux, J.L. Lisani, A.B. Petro, A nonparametric approach for histogram
segmentation, IEEE Transactions on Image Processing, 16 (2007) 253-261.
[39] B. Fadlallah, B. Chen, A. Keil, J. Príncipe, Weighted-permutation entropy: a complexity measure
for time series incorporating amplitude information, Physical Review E, 87 (2013) 022911.
[40] D. Zosso, K. Dragomiretskiy, A.L. Bertozzi, P.S. Weiss, Two-dimensional compact variational
mode decomposition, Journal of Mathematical Imaging and Vision, 58 (2017) 294-320.
[41] M. Rogalski, M. Pielach, A. Cicone, P. Zdańkowski, L. Stanaszek, K. Drela, K. Patorski, B.
Lukomska, M. Trusiak, Tailoring 2D fast iterative filtering algorithm for low-contrast optical fringe
pattern preprocessing, Optics and Lasers in Engineering, 155 (2022) 107069.
[42] J.C. Nunes, Y. Bouaoune, E. Delechelle, O. Niang, P. Bunel, Image analysis by bidimensional
empirical mode decomposition, Image and Vision Computing, 21 (2003) 1019-1026.
[43] Y. Ge, Q. Cheng, Boundary effect reduction in image filtering, Int. J. Graph. Vis. Image Process., 7
(2007) 105.
[44] A. Aricò, M. Donatelli, S. Serra-Capizzano, The anti-reflective algebra: structural and
computational analysis with application to image deblurring and denoising, Calcolo, 45 (2008) 149-175.
Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this paper.

☐ The authors declare the following financial interests/personal relationships which may be
considered as potential competing interests: